Repository navigation
Add CuPy support #1985
Description
CuPy (https://docs.cupy.dev/en/stable/index.html) is a numpy-compatible array library that uses CUDA for improved performance. What this means is that CuPy arrays live on and are computed on the GPU. It should be theoretically possible to share a CUDA buffer (or however CuPy stores the array data) with OpenGL and therefore VisPy.
This would allow for some really unique workflows where users could do very complex calculations in-GPU with CuPy and visualize them "instantly" without having to copy data back and forth between GPU and CPU memory.
Reacted by Malek Cellier, Benjamin Zaitlen, Core Francisco Park, qaz10102030, Zachary Moon, Hoàng Tùng Lâm (Linus), Tim Weber, RZS, Chester Tan, Javier Tirado Garín and 2 more
Reacted by Lorenzo Gaifas, Robert Haase, Hoàng Tùng Lâm (Linus) and Kenichi Maehashi
Reacted by Peter Sobolewski, Hoàng Tùng Lâm (Linus) and Kenichi Maehashi
Activity
djhoese commentedon Mar 17, 2021
See #1986 for some more in-depth cuda talk.
brisvag commentedon Sep 7, 2021
This would be amazing, I found myself toying with interactive tools ideas lately, and every time the bottleneck ended up being this. I'd love to be able to simply replace my
numpywithcupyand get direct access to vispy's texture buffers!djhoese commentedon Sep 7, 2021
@brisvag What type of work are you doing? Your use case sounds like the opposite of what I had in mind here. My thought is that VisPy could visualize the results of CuPy work (read only). By saying "access to vispy's texture buffers" are you saying you want to read what vispy has produced and use it in CuPy?
brisvag commentedon Sep 7, 2021
No, I think we mean the same thing, I just expressed myself poorly :) I'd like to be able to visualize quickly and in real time what happens to the data. This becomes unwieldy when working with big volumes, because every single time they are updated vispy had to make a copy of the data and transfer it to GPU memory. I tried to dive in the
glooandcupycode and understand where exactly the data gets passed to the buffer and how, but it quickly became clear that this is way over my level :)